Skip to content

Tangle

Argon proposes.

Code Video Cerebral Valley

Core trick

A geometry checker measures anamorphic billboard distortion from a simulated pedestrian view, and Argon patches the scene until it passes.

Tags

visionagents

Description

Argon proposes. The checker decides. Tangle turns “does it look right?” into a number: Gemini Argon drafts a 3D corner-billboard ad, we simulate what a pedestrian sees, and a separate geometry checker measures every error. When the illusion breaks, Argon watches the clip and patches the scene, and we measure again.
sivasuryachandran/Tangle★ 00 forksPythonCreated Oct 25, 2026 · last push Oct 25, 2026

README excerpt

# Tangle — 3D billboard ads that pass the street-view test

> Argon designs an anamorphic 3D ad for a corner screen. Tangle renders it, simulates what a
> pedestrian sees, **measures** whether the illusion holds, lets Argon repair the scene, and issues
> a certificate. **Argon proposes; an independent geometry checker decides.**

**Hackathon disclosure:** `CLAUDE.md` is a design note written before the event (before 10:30).
All code in this repo was written on Oct 25, 2026 during the hacking window.

Track: Problem Statement Two — More Dimensions. Results come from simulation, not footage of a real billboard.

## Status
Every number reported here comes from a run (`runs/`, kept locally except the committed evidence runs) and the gate log in `reports/gates.md`. Statuses: PASS / FAIL / NOT RUN.

| Gate | Status | Evidence |
|---|---|---|
| G0 Argon call works (text + inline video) | PASS | `reports/gates.md` |
| G1 identity test (assumed = true) | PASS | C2 mean 9.8e-08 deg, 32/32 ID dots within 2 reference-camera px; `reports/gates.md` |
| G2 checker mutation tests | PASS | `pytest tests/` (126 tests) |
| G3 Argon scene end to end | RAN, overall FAIL (C4, C5) | `reports/gates.md` |
| G4 live geometry change -> loop -> PASS | **PASS in the rehearsal (v4); NOT reached in the two final 6-round runs** | `g4_rehearsal` reached PASS at v4 (95 deg true corner). The later runs `g4_final` and `g5_flare` (6 rounds each, frozen limits) ended FAIL: C2 max 0.349 deg vs the 0.28 deg limit.

Team